def main():
parser = argparse.ArgumentParser()
parser.add_argument("--profile",action="store_true")
parser.add_argument("--unittest",action="store_true")
parser.add_argument("--epochs",type=int,default=10)
args = parser.parse_args()
batchsize = 64
Xshape = (batchsize, 3, 32, 32)
X = cgt.tensor4("X", fixed_shape = Xshape)
y = cgt.vector("y", fixed_shape = (batchsize,), dtype='i4')
conv1 = nn.SpatialConvolution(3, 32, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=1e-4))(X)
relu1 = nn.rectify(conv1)
pool1 = nn.max_pool_2d(relu1, kernelshape=(3,3), stride=(2,2))
conv2 = nn.SpatialConvolution(32, 32, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=0.01))(pool1)
relu2 = nn.rectify(conv2)
pool2 = nn.max_pool_2d(relu2, kernelshape=(3,3), stride=(2,2))
conv3 = nn.SpatialConvolution(32, 64, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=0.01))(pool2)
pool3 = nn.max_pool_2d(conv3, kernelshape=(3,3), stride=(2,2))
relu3 = nn.rectify(pool3)
d0,d1,d2,d3 = relu3.shape
flatlayer = relu3.reshape([d0,d1*d2*d3])
nfeats = cgt.infer_shape(flatlayer)[1]
ip1 = nn.Affine(nfeats, 10)(flatlayer)
logprobs = nn.logsoftmax(ip1)
loss = -logprobs[cgt.arange(batchsize), y].mean()
params = nn.get_parameters(loss)
updates = rmsprop_updates(loss, params, stepsize=1e-3)
train = cgt.function(inputs=[X, y], outputs=[loss], updates=updates)
if args.profile: cgt.profiler.start()
data = fetch_dataset("http://rll.berkeley.edu/cgt-data/cifar10.npz")
Xtrain = data["X_train"]
ytrain = data["y_train"]
print fmt_row(10, ["Epoch","Train NLL","Train Err","Test NLL","Test Err","Epoch Time"])
for i_epoch in xrange(args.epochs):
for start in xrange(0, Xtrain.shape[0], batchsize):
tstart = time.time()
end = start+batchsize
print train(Xtrain[start:end], ytrain[start:end]), time.time()-tstart
if start > batchsize*5: break
# elapsed = time.time() - tstart
# trainerr, trainloss = computeloss(Xtrain[:len(Xtest)], ytrain[:len(Xtest)])
# testerr, testloss = computeloss(Xtest, ytest)
# print fmt_row(10, [i_epoch, trainloss, trainerr, testloss, testerr, elapsed])
if args.profile:
cgt.profiler.print_stats()
return
if args.unittest:
break
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--profile",action="store_true")
parser.add_argument("--unittest",action="store_true")
parser.add_argument("--epochs",type=int,default=10)
args = parser.parse_args()
batchsize = 64
Xshape = (batchsize, 3, 32, 32)
X = cgt.tensor4("X", fixed_shape = Xshape)
y = cgt.vector("y", fixed_shape = (batchsize,), dtype='i4')
conv1 = nn.SpatialConvolution(3, 32, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=1e-4))(X)
relu1 = nn.rectify(conv1)
pool1 = nn.max_pool_2d(relu1, kernelshape=(3,3), stride=(2,2))
conv2 = nn.SpatialConvolution(32, 32, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=0.01))(relu1)
relu2 = nn.rectify(conv2)
pool2 = nn.max_pool_2d(relu2, kernelshape=(3,3), stride=(2,2))
conv3 = nn.SpatialConvolution(32, 64, kernelshape=(5,5), pad=(2,2),
weight_init=nn.IIDGaussian(std=0.01))(pool2)
pool3 = nn.max_pool_2d(conv3, kernelshape=(3,3), stride=(2,2))
relu3 = nn.rectify(pool3)
d0,d1,d2,d3 = relu3.shape
flatlayer = relu3.reshape([d0,d1*d2*d3])
nfeats = cgt.infer_shape(flatlayer)[1]
ip1 = nn.Affine(nfeats, 10)(flatlayer)
logprobs = nn.logsoftmax(ip1)
loss = -logprobs[cgt.arange(batchsize), y].mean()
params = nn.get_parameters(loss)
updates = rmsprop_updates(loss, params, stepsize=1e-3)
train = cgt.function(inputs=[X, y], outputs=[loss], updates=updates)
if args.profile: cgt.profiler.start()
data = np.load("/Users/joschu/Data/cifar-10-batches-py/cifar10.npz")
Xtrain = data["X_train"]
ytrain = data["y_train"]
print fmt_row(10, ["Epoch","Train NLL","Train Err","Test NLL","Test Err","Epoch Time"])
for i_epoch in xrange(args.epochs):
for start in xrange(0, Xtrain.shape[0], batchsize):
tstart = time.time()
end = start+batchsize
print train(Xtrain[start:end], ytrain[start:end]), time.time()-tstart
if start > batchsize*5: break
# elapsed = time.time() - tstart
# trainerr, trainloss = computeloss(Xtrain[:len(Xtest)], ytrain[:len(Xtest)])
# testerr, testloss = computeloss(Xtest, ytest)
# print fmt_row(10, [i_epoch, trainloss, trainerr, testloss, testerr, elapsed])
if args.profile:
cgt.profiler.print_stats()
return
if args.unittest:
break
def build_convnet_return_loss(X, y):
np.random.seed(0)
conv1 = nn.rectify(
nn.SpatialConvolution(1, 32, kernelshape=(3, 3), pad=(0, 0), weight_init=nn.IIDGaussian(std=0.1))(X)
)
pool1 = nn.max_pool_2d(conv1, kernelshape=(3, 3), stride=(2, 2))
conv2 = nn.rectify(
nn.SpatialConvolution(32, 32, kernelshape=(3, 3), pad=(0, 0), weight_init=nn.IIDGaussian(std=0.1))(pool1)
)
pool2 = nn.max_pool_2d(conv2, kernelshape=(3, 3), stride=(2, 2))
d0, d1, d2, d3 = pool2.shape
flatlayer = pool2.reshape([d0, d1 * d2 * d3])
nfeats = cgt.infer_shape(flatlayer)[1]
logprobs = nn.logsoftmax(nn.Affine(nfeats, 10)(flatlayer))
loss = -logprobs[cgt.arange(X.shape[0]), y].mean()
return loss
def convnet_model(X, w, w2, w3, w4, w_o, p_drop_conv, p_drop_hidden):
l1a = nn.rectify(nn.conv2d(X, w, kernelshape=(3, 3), pad=(1, 1)))
l1 = nn.max_pool_2d(l1a, kernelshape=(2, 2), stride=(2, 2))
l1 = nn.dropout(l1, p_drop_conv)
l2a = nn.rectify(nn.conv2d(l1, w2, kernelshape=(3, 3), pad=(1, 1)))
l2 = nn.max_pool_2d(l2a, kernelshape=(2, 2), stride=(2, 2))
l2 = nn.dropout(l2, p_drop_conv)
l3a = nn.rectify(nn.conv2d(l2, w3, kernelshape=(3, 3), pad=(1, 1)))
l3b = nn.max_pool_2d(l3a, kernelshape=(2, 2), stride=(2, 2))
batchsize, channels, rows, cols = l3b.shape
l3 = cgt.reshape(l3b, [batchsize, channels * rows * cols])
l3 = nn.dropout(l3, p_drop_conv)
l4 = nn.rectify(cgt.dot(l3, w4))
l4 = nn.dropout(l4, p_drop_hidden)
pyx = nn.softmax(cgt.dot(l4, w_o))
return pyx
def convnet_model(X, w, w2, w3, w4, w_o, p_drop_conv, p_drop_hidden):
l1a = nn.rectify(nn.conv2d(X, w, kernelshape=(3,3), pad=(1,1)))
l1 = nn.max_pool_2d(l1a, kernelshape=(2, 2), stride=(2,2))
l1 = nn.dropout(l1, p_drop_conv)
l2a = nn.rectify(nn.conv2d(l1, w2, kernelshape=(3,3), pad=(1,1)))
l2 = nn.max_pool_2d(l2a, kernelshape=(2, 2), stride=(2,2))
l2 = nn.dropout(l2, p_drop_conv)
l3a = nn.rectify(nn.conv2d(l2, w3, kernelshape=(3,3), pad=(1,1)))
l3b = nn.max_pool_2d(l3a, kernelshape=(2, 2), stride=(2,2))
batchsize,channels,rows,cols = l3b.shape
l3 = cgt.reshape(l3b, [batchsize, channels*rows*cols])
l3 = nn.dropout(l3, p_drop_conv)
l4 = nn.rectify(cgt.dot(l3, w4))
l4 = nn.dropout(l4, p_drop_hidden)
pyx = nn.softmax(cgt.dot(l4, w_o))
return pyx
def build_convnet_return_loss(X, y):
np.random.seed(0)
conv1 = nn.rectify(
nn.SpatialConvolution(1,
32,
kernelshape=(3, 3),
pad=(0, 0),
weight_init=nn.IIDGaussian(std=.1))(X))
pool1 = nn.max_pool_2d(conv1, kernelshape=(3, 3), stride=(2, 2))
conv2 = nn.rectify(
nn.SpatialConvolution(32,
32,
kernelshape=(3, 3),
pad=(0, 0),
weight_init=nn.IIDGaussian(std=.1))(pool1))
pool2 = nn.max_pool_2d(conv2, kernelshape=(3, 3), stride=(2, 2))
d0, d1, d2, d3 = pool2.shape
flatlayer = pool2.reshape([d0, d1 * d2 * d3])
nfeats = cgt.infer_shape(flatlayer)[1]
logprobs = nn.logsoftmax(nn.Affine(nfeats, 10)(flatlayer))
loss = -logprobs[cgt.arange(X.shape[0]), y].mean()
return loss
def test_pool(**kwargs):
np.random.seed(0)
x = cgt.tensor4("x", fixed_shape=(2,3,5,7))
y = max_pool_2d(x, (4,4),(0,0),(1,1))
xval = np.random.randn(2,3,5,7)
hval = np.random.randn(*cgt.infer_shape(y))
h = cgt.constant(hval)
cost = (y*h).sum()
fcost = cgt.function([x], cost)
fgrad = cgt.function([x], cgt.grad(cost, [x])[0])
from cgt.numeric_diff import numeric_grad
gnum = numeric_grad(fcost, xval)
gana = fgrad(xval)
assert np.allclose(gnum,gana)
def test_cpu_pool(**kwargs):
np.random.seed(0)
x = cgt.tensor4("x", fixed_shape=(2, 3, 5, 7))
y = max_pool_2d(x, (4, 4), (0, 0), (1, 1))
xval = np.random.randn(2, 3, 5, 7)
hval = np.random.randn(*cgt.infer_shape(y))
h = cgt.constant(hval)
cost = (y * h).sum()
fcost = cgt.function([x], cost)
fgrad = cgt.function([x], cgt.grad(cost, [x])[0])
from cgt.numeric_diff import numeric_grad
gnum = numeric_grad(fcost, xval)
gana = fgrad(xval)
assert np.allclose(gnum, gana)
def test_cpu_pool():
with cgt.scoped_update_config(precision="quad", backend="native"):
print cgt.get_precision()
ci = get_compile_info()
np.random.seed(0)
x = cgt.tensor4("x", fixed_shape=(2, 3, 5, 7))
y = max_pool_2d(x, (4, 4), (0, 0), (1, 1))
xval = np.random.randn(2, 3, 5, 7)
hval = np.random.randn(*cgt.infer_shape(y))
h = cgt.constant(hval)
cost = (y * h).sum()
fcost = cgt.function([x], cost)
fgrad = cgt.function([x], cgt.grad(cost, [x])[0])
from cgt.numeric_diff import numeric_grad
gnum = numeric_grad(fcost, xval)
gana = fgrad(xval)
assert np.allclose(gnum, gana)
def test_cpu_pool():
with cgt.scoped_update_config(precision="quad",backend="native"):
print cgt.get_precision()
ci = get_compile_info()
np.random.seed(0)
x = cgt.tensor4("x", fixed_shape=(2,3,5,7))
y = max_pool_2d(x, (4,4),(0,0),(1,1))
xval = np.random.randn(2,3,5,7)
hval = np.random.randn(*cgt.infer_shape(y))
h = cgt.constant(hval)
cost = (y*h).sum()
fcost = cgt.function([x], cost)
fgrad = cgt.function([x], cgt.grad(cost, [x])[0])
from cgt.numeric_diff import numeric_grad
gnum = numeric_grad(fcost, xval)
gana = fgrad(xval)
assert np.allclose(gnum,gana)
# reshape for convnet
Xtrainimg = Xtrain.reshape(-1, 1, 28, 28)
Xtestimg = Xtest.reshape(-1, 1, 28, 28)
# Model:
# Make it VGG-like
# VGG nets have 3x3 kernels with length 1 padding and max-pooling has all 2s.
#
# VGG is a large model so here well just do a small part of it.
X = cgt.tensor4('X', fixed_shape=(None, 1, 28, 28))
y = cgt.vector('y', dtype='i8')
conv1 = nn.rectify(
nn.SpatialConvolution(1, 32, kernelshape=(3,3), stride=(1,1), pad=(1,1), weight_init=nn.IIDGaussian(std=.1))(X)
)
pool1 = nn.max_pool_2d(conv1, kernelshape=(2,2), stride=(2,2))
conv2 = nn.rectify(
nn.SpatialConvolution(32, 32, kernelshape=(3,3), stride=(1,1), pad=(1,1), weight_init=nn.IIDGaussian(std=.1))(pool1)
)
pool2 = nn.max_pool_2d(conv2, kernelshape=(2,2), stride=(2,2))
d0, d1, d2, d3 = pool2.shape
flat = pool2.reshape([d0, d1*d2*d3])
nfeats = cgt.infer_shape(flat)[1]
probs = nn.softmax(nn.Affine(nfeats, 10)(flat))
cost = -categorical.loglik(y, probs).mean()
y_preds = cgt.argmax(probs, axis=1)
err = cgt.cast(cgt.not_equal(y, y_preds), cgt.floatX).mean()
def build_fcn_action_cond_encoder_net(input_shapes, levels=None):
x_shape, u_shape = input_shapes
x_c_dim = x_shape[0]
x1_c_dim = 16
levels = levels or [3]
levels = sorted(set(levels))
X = cgt.tensor4('X', fixed_shape=(None, ) + x_shape)
U = cgt.matrix('U', fixed_shape=(None, ) + u_shape)
# encoding
Xlevels = {}
for level in range(levels[-1] + 1):
if level == 0:
Xlevel = X
else:
if level == 1:
xlevelm1_c_dim = x_c_dim
xlevel_c_dim = x1_c_dim
else:
xlevelm1_c_dim = xlevel_c_dim
xlevel_c_dim = 2 * xlevel_c_dim
Xlevel_1 = nn.rectify(
nn.SpatialConvolution(xlevelm1_c_dim,
xlevel_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='conv%d_1' % level,
weight_init=nn.IIDGaussian(std=0.01))(
Xlevels[level - 1]))
Xlevel_2 = nn.rectify(
nn.SpatialConvolution(
xlevel_c_dim,
xlevel_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='conv%d_2' % level,
weight_init=nn.IIDGaussian(std=0.01))(Xlevel_1))
Xlevel = nn.max_pool_2d(Xlevel_2,
kernelshape=(2, 2),
pad=(0, 0),
stride=(2, 2))
Xlevels[level] = Xlevel
# bilinear
Xlevels_next_pred_0 = {}
Ylevels = OrderedDict()
Ylevels_diff_pred = OrderedDict()
for level in levels:
Xlevel = Xlevels[level]
Xlevel_diff_pred = Bilinear(input_shapes,
b=None,
axis=2,
name='bilinear%d' % level)(Xlevel, U)
Xlevels_next_pred_0[level] = Xlevel + Xlevel_diff_pred
Ylevels[level] = Xlevel.reshape(
(Xlevel.shape[0], cgt.mul_multi(Xlevel.shape[1:])))
Ylevels_diff_pred[level] = Xlevel_diff_pred.reshape(
(Xlevel_diff_pred.shape[0],
cgt.mul_multi(Xlevel_diff_pred.shape[1:])))
# decoding
Xlevels_next_pred = {}
for level in range(levels[-1] + 1)[::-1]:
if level == levels[-1]:
Xlevel_next_pred = Xlevels_next_pred_0[level]
else:
if level == 0:
xlevelm1_c_dim = x_c_dim
elif level < levels[-1] - 1:
xlevel_c_dim = xlevelm1_c_dim
xlevelm1_c_dim = xlevelm1_c_dim // 2
Xlevel_next_pred_2 = SpatialDeconvolution(
xlevel_c_dim,
xlevel_c_dim,
kernelshape=(2, 2),
pad=(0, 0),
stride=(2, 2),
name='upsample%d' % (level + 1),
weight_init=nn.IIDGaussian(std=0.01))(Xlevels_next_pred[
level +
1]) # TODO initialize with bilinear # TODO should rectify?
Xlevel_next_pred_1 = nn.rectify(
SpatialDeconvolution(
xlevel_c_dim,
xlevel_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='deconv%d_2' % (level + 1),
weight_init=nn.IIDGaussian(std=0.01))(Xlevel_next_pred_2))
nonlinearity = nn.rectify if level > 0 else cgt.tanh
Xlevel_next_pred = nonlinearity(
SpatialDeconvolution(
xlevel_c_dim,
xlevelm1_c_dim,
kernelshape=(3, 3),
pad=(1, 1),
stride=(1, 1),
name='deconv%d_1' % (level + 1),
weight_init=nn.IIDGaussian(std=0.01))(Xlevel_next_pred_1))
if level in Xlevels_next_pred_0:
coefs = nn.parameter(nn.init_array(nn.Constant(0.5), (2, )),
name='sum%d.coef' % level)
Xlevel_next_pred = coefs[0] * Xlevel_next_pred + coefs[
1] * Xlevels_next_pred_0[level]
# TODO: tanh should be after sum
Xlevels_next_pred[level] = Xlevel_next_pred
X_next_pred = Xlevels_next_pred[0]
Y = cgt.concatenate(Ylevels.values(), axis=1)
Y_diff_pred = cgt.concatenate(Ylevels_diff_pred.values(), axis=1)
X_diff = cgt.tensor4('X_diff', fixed_shape=(None, ) + x_shape)
X_next = X + X_diff
loss = ((X_next - X_next_pred)**2).mean(axis=0).sum() / 2.
net_name = 'FcnActionCondEncoderNet_levels' + ''.join(
str(level) for level in levels)
input_vars = OrderedDict([(var.name, var) for var in [X, U, X_diff]])
pred_vars = OrderedDict([('Y_diff_pred', Y_diff_pred), ('Y', Y),
('X_next_pred', X_next_pred)])
return net_name, input_vars, pred_vars, loss
def build_vgg_net(nn_input):
conv_1 = nn.convLayer(nn_input, NUM_CHANNELS, NUM_CHANNELS, kernelshape=(64, 3), pad=(1, 1))
conv_2 = nn.convLayer(conv_1, NUM_CHANNELS, NUM_CHANNELS, kernelshape=(64, 3), pad=(1, 1))
pool_1 = nn.max_pool_2d()
